vla_real_pk_remove_sharp_full_step50

Edited pi0.5 VLA checkpoint for pass_knife task — pk_remove_sharp_full arm at step 50 (early/minimal edit).

⚠️ Why this checkpoint matters

Centroid analysis revealed the LATER pk_remove_sharp_full ckpts (step 550, 600) are Goodharted: classifier P(target)=1.0 but hidden states are FAR from real {left, right} foundation manifold.

This step-50 ckpt is the only pk_full ckpt where 47% of edited-sharp frames are genuinely closer to the {left, right} foundation centroid than to the sharp centroid (vs 0% for steps 550/600 and 14% for gradgate_step400).

Hypothesis: this ckpt should reduce sharp-mode rate without the destabilization seen at later steps.

Edit ckpt	% frames closer to {L,R} centroid	Coworker eval (sharp rate)
Foundation (no edit)	0% (all in sharp manifold)	50% sharp
pk_full step 50 (this ckpt)	47% ← genuine partial shift	UNTESTED — please test!
pk_full step 550	0% (Goodharted)	60% sharp (worse)
pk_gradgate step 400	14% (partial shift)	40% sharp (better, +10pp)

Edit recipe

• steering_mode: hidden_v9_mc_softhybrid_precommit_gated
• target_subset: 0,1 (left + right)
• gating: frame_index < 999 (full trajectory, no gating — same as step 550/600 ckpts)
• γ: 0.1, β: 1.0, lr: 1e-5, batch: 32
• steps trained: 50 (very early — minimal edit before Goodharting set in)

Eval target

50-seed real-robot rollouts. Compare to:

• Foundation baseline (no edit): 50% sharp / 25% left / 25% right
• pk_full step 550 (Goodharted): 60% sharp
• pk_gradgate step 400 (best so far): 40% sharp / 40% left / 20% right

If this ckpt achieves <50% sharp, it confirms our centroid analysis: minimal editing avoids the Goodharting trap that more training induces.